From Practical Machine Learning Security to Incident Reporting by Kathrin Grosse (IBM Research, Zurich, Switzerland)

Categoria:
Seminário

Onde:
Híbrido

Local:
Sala de Seminários do DI e ZOOM

Online:
https://videoconf-colibri.zoom.us/j/92950889155?pwd=YXN6MFNwaDVxbGh4RHQ5d3N0VWhLUT09#success

Descrição:

Cybersecurity ensures the trustworthy and reliable functioning of digital systems. Currently, companies spend roughly 10% of their IT budget on cybersecurity. Thus, security becomes increasingly relevant also for emerging technologies like artificial intelligence (AI). Despite a large body of academic research, our current understanding of AI security has a critical gap. It does not cover how companies, public institutions, and non-profits use AI. This gap manifests as models are studied instead of pipelines, infeasible perturbations, or assumptions are unrealistic. This leaves us with a limited understanding of AI vulnerabilities. Meanwhile, attackers aren’t waiting. They are already exploiting these vulnerabilities, and we discuss the evidence of these real-world AI security incidents. We thus discuss a proposal for an AI security incident reporting framework to create a practical understanding of AI security threats, allowing us to take a step towards trustworthy and secure AI.